Conclusion First

“Free” can be both a weapon and a trap.

Anthropic is offering its Claude Team for one year free to startups. Lambda has raised $4 billion, bringing its valuation to $14.5 billion. The prices of AI computing resources are beginning to collapse.

Is this good news for local SMEs?

The answer is, “It depends on how you use it; the difference can be as vast as heaven and earth.”

If you jump in without thinking, you might find yourself receiving a monthly bill of several tens of thousands of yen after a year, making it hard to back out. Conversely, if you understand this structural change and act accordingly, you could see a world where operations that previously cost 3 million yen a year can be run for just 5,000 yen a month.

Which way it goes depends on the decisions made at this very moment.

Anthropic’s “One Year Free”—What’s Happening?

Anthropic has announced a program offering Claude Team (normally $30 per user per month) free for one year to startups. The target audience is companies building products on top of AI models. In other words, there is a clear intention to “increase the number of companies creating services using Claude.”

This is similar to how Google provided Google Workspace for free to startups. The strategy is to let them use it for free, create dependency, and then transition to paid plans. This is a classic strategy known as “land and expand” in the SaaS industry.

If a team of five uses it, the usual cost would be $150 per month (about 23,000 yen), totaling around 270,000 yen annually. This becomes free, which is certainly appreciated.

But the problem lies in the “second year.”

After spending a year building workflows based on Claude, creating prompts, and accumulating internal knowledge, can you really say, “Actually, it’s too expensive, so we’ll stop”? You can’t. That’s lock-in.

Viewing “Hidden Costs” in Numbers

Let’s do some specific calculations.

Year 1: Free Period

• Claude Team for 5 users: 0 yen
• API usage (usage-based billing equivalent to $50 per month): approximately 7,500 yen per month
• Annual cost: approximately 90,000 yen

Year 2 and Beyond: Regular Billing

• Claude Team for 5 users: $150 per month (about 23,000 yen)
• API usage: approximately 7,500 yen per month
• Annual cost: approximately 360,000 yen

The difference is 270,000 yen annually. While this is not an “unaffordable amount” for SMEs, it is painful as a “fixed cost that increases without thought.”

Moreover, the transition costs are troublesome. Prompts, workflows, and internal manuals created based on Claude will require 30-50% of the initial effort to be transferred to another service (like GPT-4o or Gemini). If it took 10 person-days to create, it would take 3-5 person-days to migrate. Calculating at 30,000 yen per day results in 90,000 to 150,000 yen.

In other words, starting just because it’s “free” means that in the second year, you will face an annual cost of 360,000 yen plus the inability to transition.

So What Should You Do?—Three Principles to Use “Free” as a Weapon

I don’t mean to say “don’t use it.” On the contrary, this free period should be fully utilized. However, you must adhere to the following three principles.

Principle 1: Design Without Dependency on Specific Services

Do not create prompts or workflows that are “exclusive to Claude.” By standardizing the input and output formats, operations can continue even if the underlying AI model is swapped out.

Specifically, structure your internal AI utilization through an “API interface.” By placing a thin wrapper that works with Claude, GPT-4o, or Gemini, you can reduce transition costs to one-tenth.

Principle 2: Conduct All “Value-Generating Experiments” During the Free Period

Try everything you can during the year. Automatic generation of meeting minutes, drafting quotes, automatic classification of inquiry emails, automatic translation of manuals—validate which of these works for your company while it’s free.

If you can automate 20 hours of administrative work per month, that translates to a savings of 30,000 yen per month (calculated at 1,500 yen per hour), totaling 360,000 yen annually. If this validation is completed during the free period, you can determine that it is worth paying in the second year. If not, you can stop.

Principle 3: Separate “Tasks for Humans” and “Tasks for AI”

If you try to let AI do everything, the accuracy of the AI will dictate the quality of the operations. This is the most dangerous form of lock-in. Let AI handle “drafting,” “classification,” and “summarization,” while humans make the final decisions. If this separation is established, operations will not collapse even if the AI service changes.

What Lambda’s $4 Billion Funding Means

Another piece of news: Lambda, a provider of GPU computing resources, has raised $4 billion, reaching a valuation of $14.5 billion. This move is aimed at an IPO in 2027.

Does this have anything to do with SMEs? Absolutely.

The fact that GPU cloud companies like Lambda are raising massive amounts of capital means that “the supply of AI computing resources is increasing.” When supply increases, prices drop. This is a fundamental principle of economics.

In fact, the hourly rental cost of GPUs has dramatically decreased over the past two years. The rental cost for an NVIDIA H100 was about $4 per hour at the beginning of 2023, but by 2025, it is expected to drop to around $1.50 to $2. This is less than half.

What does this mean?

Two years ago, fine-tuning a dedicated AI model for your company would have cost several hundred thousand yen per month, but now it can be done for tens of thousands of yen. Local manufacturers can adjust image recognition models with their inspection data. Local real estate companies can adjust text generation models with their property data. Such tasks have now become feasible at a “let’s give it a try” level.

Edge AI—The Option to “Not Rely on the Cloud”

Another significant development is the evolution of technology that allows AI to run on edge devices.

Research on FlashMoE has demonstrated methods for efficiently running large AI models on devices with limited memory by utilizing SSDs. Cache hit rates have improved by up to 51%, and inference speeds have increased by 2.6 times.

Translating this into the context of SMEs:

Cost Structure of Cloud AI:

• Monthly API usage fee: 10,000 to 100,000 yen
• Data must be sent to the cloud (security risks)
• Inaccessible if the internet connection is lost

Cost Structure of Edge AI:

• Initial investment: 50,000 to 300,000 yen (device purchase)
• Monthly running costs: only electricity (a few hundred to a few thousand yen)
• Data remains within the company
• Operates without the internet

For tasks like inspection in factories, inventory counting in stores, and sensor data analysis in agriculture, edge AI offers overwhelmingly better cost performance for repetitive daily tasks. It frees SMEs from monthly subscription fees, which is critically important.

The True Meaning of “Price Disruption”—The Battlefield Has Changed

Let’s summarize the discussion so far.

1. The cost of using AI tools is decreasing (Anthropic’s free offering)
2. The cost of procuring AI computing resources is decreasing (Lambda’s large funding → increase in supply → price drop)
3. The cost of hardware to run AI is decreasing (advancements in edge AI technology)

Price disruption is occurring across all three layers.

What this means is that whether or not you can use AI is no longer a differentiating factor. Both large and small companies have access to the same AI. So, what will set them apart?

It will be whether they can think about “how to solve their own on-site challenges with AI.”

This is where the seeds of reversal for local SMEs lie.

Large corporations circulate approval documents for AI implementation, hire consultants, and spend six months on PoCs (proof of concepts). SMEs can start with their president saying, “This looks interesting; let’s try it from tomorrow.” The speed of this decision-making will become the greatest asset in the AI era.

So, What Should You Do?

Just three things:

1. Try everything that is free right now. Anthropic’s free program, OpenAI’s free tier, Google’s free tier—interact with all of them. However, ensure that the design does not depend on specific services.

2. Always calculate “how much it will cost each month.” Estimate the costs after the free period before implementation. Weigh the annual costs against the labor savings. If it doesn’t add up, stop. That’s all.

3. Start small with “one task.” You don’t need to think about full company implementation. Whether it’s automating meeting minutes, drafting emails, or generating quote templates, if you can save 10 hours a month on one task, that alone is worth 15,000 yen a month.

The price disruption of AI computing resources is undoubtedly happening. However, simply “becoming cheaper” changes nothing. The key is how to utilize what has become cheaper in your own operations. Can you start that experiment today?

That is the dividing line for the future.